[Cbe-oss-dev] [PATCH 11/11] powerpc/spufs: Implement SPU affinity on top of gang scheduling

[Andre Detsch]

SPU affinity, originally implemented before we had gang
scheduling, was disabled after gang scheduling was
introduced.

This patch re-enables SPU affinity, making it fit the new scheduling
algorithm.

Signed-off-by: Andre Detsch <adetsch at br.ibm.com>
---
 arch/powerpc/platforms/cell/spufs/sched.c |   60 
+++++++++++++++++------------
 arch/powerpc/platforms/cell/spufs/spufs.h |    1 -
 2 files changed, 35 insertions(+), 26 deletions(-)

diff --git a/arch/powerpc/platforms/cell/spufs/sched.c 
b/arch/powerpc/platforms/cell/spufs/sched.c
index 8326034..c34e53f 100644
--- a/arch/powerpc/platforms/cell/spufs/sched.c
+++ b/arch/powerpc/platforms/cell/spufs/sched.c
@@ -443,31 +443,20 @@ static struct spu *ctx_location(struct spu *ref, int 
offset, int node)
 	return spu;
 }
 
-/*
- * affinity_check is called each time a context is going to be scheduled.
- * It returns the spu ptr on which the context must run.
- */
-static int has_affinity(struct spu_gang *gang)
+static void set_affinity(struct spu_gang *gang)
 {
-	if (list_empty(&gang->aff_list_head))
-		return 0;
-
-	/*
-	 * TODO: fix SPU Affinity to work with gang scheduling.
-	 */
-
-	if (atomic_read(&gang->aff_sched_count) == 0)
-		gang->aff_ref_spu = NULL;
+	BUG_ON(list_empty(&gang->aff_list_head));
 
-	if (!gang->aff_ref_spu) {
-		if (!(gang->aff_flags & AFF_MERGED))
-			aff_merge_remaining_ctxs(gang);
-		if (!(gang->aff_flags & AFF_OFFSETS_SET))
-			aff_set_offsets(gang);
-		aff_set_ref_point_location(gang);
-	}
+	if (!(gang->aff_flags & AFF_MERGED))
+		aff_merge_remaining_ctxs(gang);
+	if (!(gang->aff_flags & AFF_OFFSETS_SET))
+		aff_set_offsets(gang);
+	aff_set_ref_point_location(gang);
+}
 
-	return gang->aff_ref_spu != NULL;
+static int has_affinity(struct spu_gang *gang)
+{
+	return !list_empty(&gang->aff_list_head);
 }
 
 /**
@@ -486,9 +475,6 @@ static void spu_unbind_context(struct spu *spu, struct 
spu_context *ctx)
  	if (spu->ctx->flags & SPU_CREATE_NOSCHED)
 		atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
 
-	if (ctx->gang)
-		atomic_dec_if_positive(&ctx->gang->aff_sched_count);
-
 	spu_switch_notify(spu, NULL);
 	spu_unmap_mappings(ctx);
 	spu_save(&ctx->csa, spu);
@@ -582,6 +568,15 @@ static struct spu *spu_bind(struct spu_gang *gang,
 		if (!node_allowed(gang, node))
 			continue;
 
+		if (has_affinity(gang)) {
+			spin_lock(&cbe_spu_info[node].list_lock);
+			spu = ctx_location(gang->aff_ref_spu, ctx->aff_offset,
+					   node);
+			if (spu && spu->alloc_state == SPU_FREE)
+				goto found;
+			spin_unlock(&cbe_spu_info[node].list_lock);
+		}
+
 		spin_lock(&cbe_spu_info[node].list_lock);
 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 			if ((spu->alloc_state == SPU_FREE) &&
@@ -608,6 +603,9 @@ static void __spu_schedule(struct spu_gang *gang, int 
node_chosen)
 
 	spu_del_from_rq(gang);
 
+	if (has_affinity(gang))
+		set_affinity(gang);
+
 	list_for_each_entry(ctx, &gang->list, gang_list) {
 		mutex_lock(&ctx->state_mutex);
 		BUG_ON(ctx->spu);
@@ -657,6 +655,18 @@ static int spu_get_idle(struct spu_gang *gang, int node)
 	spu_context_nospu_trace(spu_get_idle__enter, gang);
 
 	/* TO DO: SPU affinity scheduling. */
+#if 0
+	if (has_affinity(gang)) {
+		aff_ref_spu = ctx->gang->aff_ref_spu;
+		node = aff_ref_spu->node;
+
+		mutex_lock(&cbe_spu_info[node].list_mutex);
+		spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
+		if (spu && spu->alloc_state == SPU_FREE)
+			goto found;
+		mutex_unlock(&cbe_spu_info[node].list_mutex);
+	}
+#endif
 
 	mode = SPU_RESERVE;
 
diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h 
b/arch/powerpc/platforms/cell/spufs/spufs.h
index 6afc514..907baf9 100644
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -178,7 +178,6 @@ struct spu_gang {
 	struct mutex aff_mutex;
 	int aff_flags;
 	struct spu *aff_ref_spu;
-	atomic_t aff_sched_count;
 
 	/* spu scheduler statistics for zombie ctxts */
 	struct {
-- 
1.5.4.1